A New Subspace Identification Approach Based on Principal
Component Analysis and Noise Estimation
Ping Wu,*
,†
HaiPeng Pan,
†
Jia Ren,
†
and Chunjie Yang
‡
†
Faculty of Mechanical Engineering & Automation, Zhejiang Sci-Tech University, Hangzhou 310018, Zhejiang, China
‡
State Key Laboratory of Industrial Control Technology, Zhejiang University, Hangzhou 310027, Zhejiang, China
ABSTRACT: In this paper, a new subspace identification approach based on principal component analysis (PCA) and noise
estimation is developed for multivariable dynamic process modeling. In contrast to typical subspace identifi cation methods based
on standard PCA with instrumental variables, the noise term is first estimated and naturally eliminated in the proposed approach,
and then a PCA procedure is used to determine system observability subspace and extract system matrices A, B, C, and D from
the estimated observability subspace. For comparison with other typical subspace identification methods based on PCA,
numerical simulation and activated sludge process benchmark modeling are included to demonstrate the superiority of the
proposed approach and reveal the probable reason for unsatisfied B and D estimations derived by some subspace identification
methods based on PCA.
1. INTRODUCTION
Subspace identification methods have gained great interest from
both academy and industry in the past decades.
1−9
Compared
with prediction error methods (PEM),
10
the most attractive
merits of subspace identifi cation methods (SIM) are ease-of-
extension for multi-input multi-output (MIMO) systems and
simple-of-computation without nonlinear optimization through
robust numerical linear algebra tools such as Singular Value
Decomposition (SVD) and QR factorization.
11
The general SIM procedures for open loop are as follows:
12
(1) Identification of the extended observability matrix, block
triangular Toeplitz matrix, or realized state sequence through
projection and model reduction processes. (2) Estimation of the
system parameters from estimate of extended observability
matrix and block triangular Toeplitz matrix by regression.
Recently, the works of analyzing statistical properties of several
SIMs have been carried out.
13−16
The dynamic principal component analysis (DPCA) method
has been used for system identification.
17
Subspace identification
based on the indirect DPCA approach is provided for modeling
the errors-in-variables (EIV) process.
18
However, in ref 19, the
authors pointed out that there are some limitations for SIM based
on DPCA, such as all variables must have identical noise variance
for deriving consistent estimate. A new subspace identification
approach called SIMPCA or SMIPCA that employs the PCA
procedure to determine the extended observability subspace and
block triangular Toeplitz matrix is proposed in the literature.
19
In
the SIMPCA algorithm, standard PCA is modified with an
instrumental variable consisting of past inputs and past outputs.
The modified PCA is adopted to determine the observability
matrix and the Toeplitz matrix from the parity space equation,
and then system matrices A, B, C, and D are extracted through the
estimated observability matri x and Toeplitz matrix. Later ,
another subspace identification-based PCA denoted as SIMP-
CA-Wc is developed by introducing column weighting for
solving the closed-loop identification problem.
20
The above-
mentioned algorithms SIMPC A and SIMPCA-Wc can be
formulated in a framework that includes different instrumental
variables. In ref 21, the authors revised subspace identification
methods based on PCA and developed a new closed-loop
subspace identification algorithm (SOPIM) by extending the
SIMPCA algorithm. Recently, the application of SIMPCA has
been generalized from fault detection
22
to fault-tolerant con-
trol.
23
The SIMPCA approach requires instrumental variables to
be highly correlated with future inputs and future outputs and
uncorrelated to the noise term. However, the condition is not
always stringently satisfied in the industrial process. The errors
caused by adopting instrumental variables may be vital for
estimating zeros of the identified process with finite data length.
In this paper, we present a new subspace identification method
based on basic SIMPCA. The proposed method does not use
instrumental variables to remove the noise term. We first
estimate the noise term and then the estimated noise sequences
are treated as known variables similar to IEM algorithms.
24
Then,
the subsequent PCA procedure is implemented to determine the
system observability subspace and Toeplitz matrix. The
proposed method in this paper will be denoted as SIMPCA-E,
which means the noise term is estimated and eliminated.
The SIMPCA-E method can avoid the usage of instrumental
variables. Simulations are conducted to demonstrate the
performance of the SIMPCA-E method with comparison to
other SIMs.
This paper is focused on subspace identification based on PCA
for open loop and organized as follows. In Section 2, the problem
formulation and assumptions are introduced. Section 3 gives a
brief review of the basic SIMPCA algorithm and presents the
SIMPCA-E method. Simulations are given on the SIMPCA-E
method compared with other methods in Section 4. Conclusions
are provided in the final section.
Received: December 17, 2014
Revised: March 16, 2015
Accepted: April 19, 2015
Published: April 19, 2015
Article
pubs.acs.org/IECR
© 2015 American Chemical Society 5106 DOI: 10.1021/ie504824a
Ind. Eng. Chem. Res. 2015, 54, 5106−5114